danube

A 1,000-utterance Bambara speech sample — 0.87 hours of 16 kHz audio with transcripts and per-utterance speaker labels. Small enough to be a working sample rather than a training corpus.

Load

from datasets import load_dataset

ds = load_dataset("djelia/danube", split="train")
print(ds[0]["text"], ds[0]["speaker_id"], ds[0]["duration"])

One config and one split.

Config	Split	Rows	Audio
default	train	1,000	0.865 h

Fields

Field	Type	Notes
audio	Audio	16 kHz WAV
text	string	Bambara transcript, standard orthography
duration	float64	Seconds; mean 3.11, min 0.47, max 36.46
speaker_id	int32	55 distinct values in [-1, 67]

Notes

Text is in standard Bambara orthography using ɛ ɔ ɲ ŋ — 843 of the 1,000 rows contain at least one. French code-switching, editorial brackets ([a ye]) and tone marks (Màa y'í tìle kɛ́) all appear, so the transcripts come from more than one annotation pipeline.

speaker_id = -1 is a sentinel for unassigned speaker on 16 rows. Filter it out for speaker-conditioned work:

known = ds.filter(lambda row: row["speaker_id"] >= 0)   # 984 rows

djelia/Denube-final is the preprocessed, ASCII-folded form of the same corpus at larger scale. This repo is the one that still has playable waveforms and correct Bambara spelling.